Вы находитесь на странице: 1из 3

Pre- and post-ictal brain activity characterization

using combined source decomposition and


connectivity estimation in epileptic children
Ivan Kotiuchyi Ivan Seleznov, Anton Popov Volodymyr Kharytonov
Biomedical Engineering Department, Electronic Engineering Department, TMO ”Psychiatry”
National Technical University of Ukraine National Technical University of Ukraine Kyiv, Ukraine
”Igor Sikorsky Kyiv Polytechnic Institute”, ”Igor Sikorsky Kyiv Polytechnic Institute”, vkharytonov69@gmail.com
R&D Engineering, Ciklum, R&D Engineering, Ciklum,
Kyiv, Ukraine Kyiv, Ukraine
ivanellokot@gmail.com ivan.seleznov1@gmail.com
popov.kpi@gmail.com

Abstract—In this research, the study of the coherence between is a need to create new methods for evaluating functional
common spatial patterns between EEG segments well before the and effective relationships during epilepsy. This connectivity
seizure, pre-ictal and post-ictal was performed. EEG recordings characterization may be obtained through the construction of
were obtained from 4 subjects with generalized seizures and 12
subjects with focal epilepsy seizures. According to the obtained autoregressive mathematical models of brain activity signals,
results of functional connectivity, it was found that the class of such as EEG, and applying signal processing tools in a spectral
EEG may be predicted from another class with the time delay of domain [4], [13]–[15]. In this work, we study the coherence
256 ms. Our results, also show that coherence between patterns between the common spatial patterns for background EEG
also differs between two types of seizures. In the case of focal and EEG for pre- and post-ictal intervals. This is applied to
seizures, the coherence, in general, has the same values for brain
activity during baseline and pre-ictal periods. For generalized the EEG from epileptic children with generalized and focal
seizures, the coherence during the pre-ictal period is larger than seizures.
during the baseline, for all patterns in the whole frequency range.

Index Terms—epilepsy, epileptic seizures, EEG, brain connec- II. S OURCE DECOMPOSITION AND CONNECTIVITY
tivity, common spatial patterns, VAR model, EEG, ICA CHARACTERIZATION

I. I NTRODUCTION A. Common Spatial Patterns


The human brain is a large-scale network in which activ- To obtain the patterns to which the EEG can be decom-
ity depends on interactions between different brain regions. posed, the Common Spatial Patterns (CSP) [5], [16]–[18]
Characterization of brain electrical activity, particularly syn- were extracted. CSP is basically the filter algorithm which
chronous functional and causal effective connectivity between provides us with new decomposed features calculated from
brain regions, is used to determine if they are altered in a wide covariance matrices of raw EEG signals, compared for fol-
variety of neurological disorders, such as epilepsy [1], [2]. lowing classes: ”baseline”–”pre-ictal”, ”baseline”–”post-ictal”,
Functional connectivity is defined as the temporal correla- ”pre-ictal”–”post-ictal”. CSP provides an appropriate spatial
tion between spatially remote brain regions. In turn, effective resolution of our high-dimensional EEG data. We reduce data
connectivity is defined as the influence of one brain region dimensionality by linear transformation and separating EEG
exerts over another. Overall connectivity describes how brain signals into additive subcomponents which hold the most dis-
regions interact and engage each other [1], [3]. Research of criminative information for each of the compared classes. The
studies involving epilepsy and other physiologic states con- duration of each class EEG data windows was set to 5 seconds
nectivity estimation was conducted. Most of studies designed and in each class, we suppose to observe different sources’
to a better understanding of the occurrence and spreading of activations. During the initiation phase of our research, we set
an epileptic seizure and estimation of brain activity during the number of dimensionality reduced features extracted from
different brain strains [4], [5], [8]–[11]. raw EEG signals with the CSP algorithm to 4. Each of these
Traditional approaches to the study of functional brain 4 spatial patterns represents the activity of cortical sources.
connectivity in epilepsy do not fully solve the problem We compare the activity between sources of the first EEG
of unification and standardization of views concerning the signal class and activity extracted from the same sources of
epileptic seizures occurrence and spreading. Therefore, there the corresponding second compared class.
B. MVAR modeling
The next step after applying CSP algorithm is to obtain
an MVAR model. We assessed the quality of MVAR models
by using the Schwartz–Bayes information criterion [19]. The
MVAR model order was set to 64. As far as the sampling
frequency of our signals is 250 Hz, this implies that MVAR
modeling of common spatial patterns is performing taking into
account the information which they contain for the time delay
of 256 ms.

C. Decomposing the VAR model residuals with ICA


We fit the MVAR models to the EEG data by decomposing
models’ residuals with ICA method and iteratively correcting
models’ coefficients. After performing the ICA method we Fig. 1. ”Baseline”–”pre-ictal” coherence between spatial patterns for gener-
get the estimated EEG source activities and obtain functional alized seizures.
connectivity from various spectral measures between spatial
sources.
EEG in each segment is considered stationary, leading to the
D. MVAR model residuals’ whiteness testing assumptions that the sources are stationary as well. In this
Additionally, the residuals were also tested for whiteness work, we observe the EEG signals frequency range of 0-45
using the Li-McLeod as Portmanteau test statistic Q [5], [20]. Hz.
For obtained models in the present study this test shows no IV. E XPERIMENTAL RESULTS
correlated residuals and the hypothesis of white residuals was
The examples of results of coherence between four spatial
not rejected. This indicates that the VAR model describes the
components are presented for generalized Fig. 1, and focal
data properly.
Fig. 2, seizures. For every pair of EEG segments (baseline -
To summarize, we performed blind source decomposition, pre-ictal, baseline - post-ictal, and pre-ictal - post-ictal), the
connectivity estimation and visualization of results using SCoT common spatial patterns were calculated and plotted along
– a Python toolbox for EEG source connectivity estimation, the main diagonal. The intensity of color corresponds to the
described in [5], [13]. distribution of the patterns values over the head surface, and
the range is normalized for four patterns in each case, light
III. EEG DATA color corresponds to high values.
EEG was recorded from 4 subjects with generalized seizures The plots show coherence spectra for baseline and pre-
and 12 subjects with focal epilepsy. In the focal group, some ictal common spatial patterns which represent the activity
subjects had focal initiation of a seizure but focal initiation of cortical sources. This connectivity measure doesnt keep
of pathologic activity. The information about participating the information about the direction of functional connectivity,
subjects is given in Table 1. just estimates the extent to which one class of EEG may be
predicted from another class with the time delay of 256 ms
TABLE I (64 model order). Moreover, coherence represents the linear
I NFORMATION ABOUT PARTICIPATING SUBJECTS relationships between activation sources of baseline (blue) and
Female Male: Number pre-ictal (green) EEG classes.
Group number, number) of The common spatial patterns for the cases of focal and
(age range) (age range) seizures generalized seizure have a different form. For the case of focal
Generalized 4, (2.8–15 y.o.) 0 18
Focal 4, (3–25 y.o.) 8, (0.5–9 y.o.) 45 seizures, CSPs are more uniform and concentrated around
the center of the head. Three CSPs have negative values in
the center and positive at the circumference. For generalized
In the present study, the two-steps approach to connectivity seizures, CSP is non-uniform: first CSP is high in the center
analysis is adopted [5]: sources of activity and connectivity and at the circumference, second CSP has low values at the
itself are estimated from the different data segments. Hence, circumference and in the area under Fz. Third CSP has low
we divided the multichannel EEG into three segments: values in the almost entire right hemisphere and under Fz and
• Baseline, starting 35 sec before a seizure and ending 30 F3 electrodes, and fourth CSP has two high values around F4
sec before a seizure. and Pz.
• Pre-ictal, staring 5 sec before a seizure and ending at the Coherence between patterns also differs between two types
moment of seizure start. of seizures. For focal seizures, the coherence, in general, has
• Postictal, starting immediately after seizure stop and the same values for brain activity during baseline and pre-
ending after 5 sec. ictal periods. There are frequency ranges where the difference
[6] L. Faes, and G. Nollo, “Multivariate Frequency Domain Analysis of
Causal Interactions in Physiological Time Series”, Biomedical Engineer-
ing, Trends in Electronics, Communications and Software, vol. 8, pp.
403–428, 2011.
[7] P. Van Mierlo, S. Assecondi, S. Staelens, “Changes in connectivity
patterns in the kainate model of epilepsy”, Biomedical Engineering,
Trends in Electronics, Communications and Software, vol. 22, pp. 360–
363, 2009.
[8] M. Di Bono, K. Priftis, C. Umilta, “Bridging the Gap between Brain
Activity and Cognition: Beyond the Different Tales of fMRI Data
Analysis”, Front Neuroscience, vol. 11, pp. 1–4, 2017.
[9] M. Cole, S. Pathak, W. Schneider, “Identifying the brain’s most globally
connected regions”, NeuroImage, vol. 10, pp. 3132–3148, 2010.
[10] C. Beckmann, M. DeLuca, J. Devlin, “Investigations into resting-
state connectivity using independent component analys”, Philosophical
Transactions of the Royal Society of London, vol. 360, pp. 1001–1013,
2005.
[11] M. Beckmann, H. Johansen-Berg, M. Rushworth, “Connectivity-based
Fig. 2. ”Baseline”–”pre-ictal” coherence between spatial patterns for focal parcellation of human cingulate cortex and its relation to functional
seizures. specialization”, The Journal of Neuroscience, vol. 29, pp. 1175–1190.,
2009.
[12] G. Gomez-Herrero and M. Atienza and K. Egiazarian and J. L. Cantero,
“Measuring directional coupling between EEG sources”, NeuroImage,
is present, e.g. for first and second pattern, coherence in the vol. 43(3), pp. 497–508, 2008.
range from 0 to 15 Hz is larger during pre-ictal, while the [13] M. Billinger, C. Brunner, G. R. Mller-Putz, “Online visualization of
brain connectivity”, Journal of Neuroscience Methods, vol. 256, pp.
coherence between first and last CSPs in the range from 25 106–116, 2015.
to 40 Hz is larger during baseline segment. [14] G. Purvi, J. Raviraj, S. Mriganka and M. Hema, “A Common Spatial
For generalized seizures, the coherence during the pre-ictal Pattern Approach for Classification of Mental Counting and Motor
Execution EEG”, 10th International Conference, Allahabad, India, Con-
period is larger than during the baseline, for all patterns in ference proceedings, pp. 26-35, December 7–9, 2018.
the whole frequency range. The most prominent difference [15] A. Schlogl and G. Supp, “Analyzing event-related EEG data with
is between third and fourth patterns from 0 to 30 Hz, their multivariate autoregressive parameters”, Progress in Brain Research, vol.
159, pp. 135–147, 2006.
coherence during baseline is approximately three times larger. [16] A. J. Bell and T. J. Sejnowski, “An Information-Maximization Approach
to Blind Separation and Blind Deconvolution”, Neural Computation, vol.
V. C ONCLUSIONS 7, pp. 1129–1159, 1995.
[17] T.W. Lee, M. Girolami and T. Sejnowski, “Independent component
Study of the coherence between common spatial patterns analysis using an extended Infomax algorithm for mixed subGaussian
between EEG segments well before the seizure, pre-ictal, and and superGaussian sources”, Neural Computation, vol. 11, pp. 417–441,
1999.
post-ictal allow us to understand the connectivity of brain [18] M. Amama, Z. Rida, B. Rushda, S. Maryam and K. Awais, “Classifica-
areas. For further analysis, the study of activations of the CSPs tion of multi-class motor imagery EEG using four band common spatial
will be performed, which will shed light on the difference pattern”, Annual International Conference of the IEEE Engineering in
Medicine and Biology Society. IEEE Engineering in Medicine and
between activations with time and for various conditions. Biology Society. Conference proceedings, pp. 1034–1037, 2017.
Moreover, such measures as partial coherence, partially [19] C. Porcaro and F. Zappasodi and P. Maria Rossini and F. Tecchio,
directed coherence, generalized partial directed coherence, “Choice of multivariate autoregressive model order affecting real net-
work functional connectivity estimate”, Clinical Neurophysiology, vol.
directed transfer function, full frequency directed transfer 120, pp. 436–448, 2009.
function, as well as generalized directed transfer function [20] E. Mahdi, “Portmanteau test statistics for seasonal serial correlation in
would we calculated and further analyzed. time series models”, SpringerPlus, vol. 5, pp. 1–13, 2016.
Also, the validity of the CSP with respect to the localization
of the epileptic focus should be conducted.

R EFERENCES
[1] K. Friston, “Functional and effective connectivity: a review,” Brain, vol.
1, pp. 13–36, 2011.
[2] J. Engel Jr., P. M. Thompson, J. M. Sterna, R. J. Stabaa, A. Bragina, and
I. Modya “Connectomics and epilepsy”, Current opinion in neurology
vol. 26,2 pp. 186–194, 2013.
[3] D. Goldenberg and A. Galvn, J. M. Sterna, R. J. Stabaa, A. Bragina,
and I. Modya “The use of functional and effective connectivity tech-
niques to understand the developing brain”, Developmental Cognitive
Neuroscience vol. 12 pp. 155 – 164, 2015.
[4] L. Faes, S. Erla, and G. Nollo, “Measuring Connectivity in Linear Mul-
tivariate Processes: Definitions, Interpretation, and Practical Analysis”,
Computational and Mathematical Methods in Medicine, vol. 2012, pp.
1–18, 2012.
[5] M. Billinger, and C. Brunner, and G. Mller-Putz, “SCoT: a Python
toolbox for EEG source connectivity”, Frontiers in Neuroinformatics,
vol. 8, pp. 1–22, 2014.

Вам также может понравиться